Introduction
Autonomous vehicle (AV) research has traditionally prioritized safety and reliability in structured urban environments. However, the Indy Autonomous Challenge (IAC) shifts the focus to the "edge cases" of vehicle physics. This competition requires full-scale Dallara AV-21 race cars to compete head-to-head at speeds exceeding 220 km/h (136 mph). Team KAIST developed a resilient, full-stack autonomous racing system designed to handle the extreme dynamics, real-time constraints, and adversarial interactions inherent in high-speed racing.
System Architecture and Methods
The KAIST autonomy stack is built on three core principles: dependability, evolvability, and performance.
System Status Manager (SSM)
The "brain" of the system's resilience is the System Status Manager. The SSM monitors the health of every software node in real-time. A critical feature is its ability to detect failures in primary GNSS-based localization using Mahalanobis distance. If the system becomes "unconfident" in its position, the SSM automatically triggers a degraded mode or an emergency perception-based controller to ensure vehicle safety.
Multi-modal Perception and Overtaking
To track opponents, the team utilized a fusion of LiDAR and Radar data processed through an Interacting Multiple Model (IMM) Unscented Kalman Filter (UKF) with Probabilistic Data Association Filter (PDAF). Notably, cameras were excluded from primary detection due to the difficulty of obtaining sufficient high-speed training data to ensure reliability.
For overtaking decisions, the team adapted Rendezvous Guidance (RG), a technique originally used for space docking. By projecting a "virtual target" ahead of an opponent, the system mathematically determines if a safe overtaking trajectory is physically feasible within the vehicle's current acceleration limits.
High-Frequency Control
To maintain stability at extreme speeds, the system employs an explicit Linear-Quadratic Regulator (LQR). To achieve a 100Hz control frequency, the Algebraic Riccati Equations were solved offline. The resulting gains are stored in a look-up table indexed by velocity, allowing for rapid, stable control without the computational overhead of online optimization.
Experimental Results
The system was validated using competition logs from the Indianapolis Motor Speedway (IMS) and Las Vegas Motor Speedway (LVMS), using rival team GPS logs as ground truth for perception accuracy.
- Performance: The vehicle achieved a maximum solo speed of 227.2 km/h and maintained stability during head-to-head racing at 212.5 km/h.
- Precision: Mean tracking error was limited to 0.3 meters at speeds over 150 km/h, though deviations reached 1.15m during aggressive cornering.
- Latency: The control loop consistently operated under the 10ms target. LiDAR clustering was identified as the main computational bottleneck, averaging 10.5ms.
- Reliability: Team KAIST was one of only three teams to complete the IAC events without a single accident, demonstrating the effectiveness of their hierarchical fail-safe architecture.
Discussion and Future Work
While successful, the system encountered "phantom braking" due to false positive detections of track walls. Future improvements will focus on more sophisticated geofencing and noise filtering. Additionally, the current overtaking logic assumes opponents maintain constant speed and lanes. Future iterations will incorporate game-theoretic models to predict and counter adversarial maneuvers in multi-car scenarios.
The team also noted a consistent "positive cross-track error" during high-speed turns, attributed to the asymmetrical "oval-track" suspension (camber) of the Dallara AV-21. This highlight the importance of high-fidelity hardware modeling in software-driven autonomous systems operating at the limit of performance.
Summary updated